bio2C: A Manually Curated Biomedical Text2Cypher Benchmark

bio2C is a manually curated biomedical Text-to-Cypher benchmark containing 2,500 natural language–Cypher pairs grounded in RNA-KG. The benchmark covers five query categories of increasing complexity, ranging from node-level retrieval and multi-hop graph traversals to aggregation, ranking, filtering, and advanced Cypher constructs. This Zenodo record contains: bio2C.zip: the complete bio2C benchmark; MedT2C.zip: a deduplicated version of the original MedT2C dataset, along with the manually curated subset; llama3_lora_mirnakgt2c.zip: the LoRA adapter weights for the fine-tuned LLaMA-3.1-8B model; miRNAKGt2c.dump: a Neo4j dump of the RNA-KG view used for creating bio2C and for running experiments.